A king-of-the-hill subnet for small open language models. Miners train a Quasar 3B mixture-of-experts model, publish the pinned weights to Hugging Face, and commit the revision on-chain. Validators run paired-KL duels and a multi-axis composite evaluator, then hand all weight to the single best model standing.
//What is Quasar
Quasar is a Bittensor subnet run by SILX Labs that turns small-model training into a continuous, on-chain tournament. Miners submit their own fine-tunes of the Quasar 3B foundation model as public Hugging Face repositories. Validators evaluate every valid submission and the strongest model wins the crown, which collects the subnet's full validator weight until something better arrives.
The simple version: It is a leaderboard you cannot game. Miners train models and pin them publicly. Validators run the same tests on every entry. The current best model is the king. Beat the king and you become the king.
Centralized equivalent: Open LLM Leaderboard or LMSYS Chatbot Arena, except entries are committed to a blockchain, evaluation runs continuously, and the winning model captures real rewards instead of just a ranking position.
How it works:
Miners train Quasar-compatible models within a fixed parameter cap, host them as pinned public Hugging Face revisions, and submit one repo per registered .
Validators run pre-checks (architecture, tokenizer, safetensors, dedup), score each model with paired-KL duels plus a composite evaluator covering capability, conversational quality, generation discipline, and robustness, then assign one-hot weights to the current king.
Why This Matters
Keep exploring
Other research from the same neighborhood of the network.
The problem it solves: Open-model benchmarks reward a single submission once, then stop. Quasar reruns the contest forever, with cryptographically pinned entries and live incentives, so improvements show up immediately in chain rewards.
The opportunity: Small-model quality is one of the most contested fronts in AI. A 3B model that can credibly hold its ground on reasoning, code, instruction following, and long context is commercially useful and cheap to serve.
The Bittensor advantage: The crown gets the entire validator weight, so the incentive to dethrone is direct and unambiguous. There is no gradient of partial credit to coast on.
Traction signals: 78 commits since the late-April 2026 rewrite, daily pushes from the lead engineer, and a published base checkpoint at silx-ai/Quasar-3B-A1B-Preview. The team also ships a separate live dashboard for the king and the chain-revealed weight target.
//Full Analysis
Category: Model Fine-Tuning | Centralized Competitor: Open LLM Leaderboard, LMSYS Arena, Hugging Face model hub competitions.
Quasar sits in the small-model contest space that has become the most active corner of open AI. The bet is that a 3B total, roughly 1B active mixture-of-experts trained against a strong teacher and scored on a hard composite of capability and robustness probes can produce a model worth running. The subnet pairs that bet with a winner-take-all incentive structure that is unusually sharp by Bittensor standards.
Mechanism:
Miners build on a fixed base. The canonical architecture is defined by silx-ai/Quasar-3B-A1B-Preview, with vocab_size 248320 and a parameter cap enforced by validators. Submissions must be loadable with trust_remote_code, must avoid quantized formats (GPTQ, AWQ, GGUF, FP8), and must remain unchanged after the committed revision. Identical weight reuse is disqualified as COPY, deleted or modified repos as REMOVED.
Evaluation is two-layered. Validators run a composite evaluator that scores distribution match against a teacher (default Qwen/Qwen3.5-4B), capability axes (math, code, reasoning, instruction following, tool use, long context, robustness), conversational quality, and generation discipline that penalizes models which ramble, loop, or never answer. Prompts are block-seeded and rewritten so miners cannot overfit to a static answer key. The composite produces both a worst-axis score and a weighted aggregate.
The second layer is paired-KL duels between the challenger and the current king. Dethroning requires both a valid paired-KL win and a composite quality pass. Composite alone cannot override a direct head-to-head regression, and KL alone cannot crown a challenger. The winner of that two-test gate gets validator weight 1.0 on-chain. Everyone else gets 0.0, with a no-winner fallback UID configured so vTrust stays aligned when no submission passes safely.
Onchain reality matches the design at the validator side and complicates the picture at the market side. The repo has had a push every day or two through early June 2026, with recent commits stabilizing the crown quality gate, the validator round guards, and a 90/10 default safety split. Active miner count from TaoSwap reads 1, which is consistent with a king-of-the-hill design where one model holds the throne and validators set one-hot weights to that single UID rather than spreading weight across many active workers.
The market side is colder. Price sits at 0.00862 TAO, down 9.6% on the day, 20.3% on the week, and 37.4% on the month. is 48,880 TAO with about 12,340 TAO of root in the pool. Net 7-day flow is negative at roughly 787 TAO outflow, and the subnet's current share is 0% under , which is the default for any subnet with negative over the window. is 0.157, so most of the pool is organic rather than protocol subsidy. Top-100 Gini sits at 0.574 with HHI 0.021, on the broad side of the distribution chart.
Quasar previously ran a different mechanism on this netuid (a LongBench-style long-context evaluation design), and the on-chain owner address has changed since the previous public write-up. The current GitHub repository, dashboard, and base checkpoint all point to the king-of-the-hill foundation model design described above. The TAO.app about page still describes the older mechanism, so treat that source as stale and the GitHub README and live dashboard as authoritative.
//Risk Factors
These factors move fast; captured at publishing date
Bus factor: Two GitHub contributors, with ahmed3520 doing essentially all of the recent work. A single key engineer carries the validator codebase.
Outflows under a 0% emission share: Taoflow gives zero emissions to subnets running negative net flows. A 7-day outflow near 787 TAO keeps the subnet in that bucket until staking flows turn positive.
Mechanism continuity: The current design is a fresh build on a re-registered netuid with a new owner address. Buyers comparing this Quasar to the version that traded earlier in the year are looking at a different team and a different product.
One-hot weight risk: King-of-the-hill means a single model captures all validator weight. A buggy crown, a misconfigured threshold, or a flawed evaluation pass can concentrate emissions on the wrong model until the next round corrects it.
Competition: Small-model leaderboards are crowded, and centralized venues have far more eyes on them. Quasar needs the cryptographic-commit angle and the incentive sharpness to translate into actual model quality that someone wants to run.
